Abstract: Existing YOLO-series-based detection methods for personal protective equipment wearing often suffer from missed detections and false detections in complex backgrounds, occlusions, and small object scenarios. When RT-DETR-based models are applied to personal protective equipment wearing detection in underground coal mine working-at-height operations, they face issues such as excessive model parameters and inaccurate bounding box localization caused by human pose variations, limb occlusion, and uneven illumination. To address these problems, an improved RT-DETR-based detection model for personal protective equipment wearing in underground coal mine working-at-height operations was proposed. The ResNet backbone in RT-DETR was replaced with the lightweight ShuffleNetv2 network, significantly reducing the number of model parameters while maintaining detection accuracy, enabling deployment on resource-constrained underground edge devices. A Focaler-MPDIoU loss function was introduced, combining fine-grained constraints on bounding box corner deviations and an adaptive focusing mechanism, enhancing detection accuracy and bounding box regression quality, and improving localization performance for occluded, pose-varying, and small targets. Quantization-aware training was further adopted to compress the model size and improve its adaptability for deployment on underground edge devices. Experimental results showed that the proposed model achieved an accuracy of 97.2% on a self-built dataset, with a parameter size of 14.5×10⁶, representing a 65.4% reduction compared with the original model and greatly improving the lightweight level of the model. Compared with the mainstream lightweight models YOLOv5n, YOLOv8n, and YOLOv11n, mAP@50 was improved by 4.1%, 4.0%, and 3.2%, respectively, demonstrating superior detection accuracy.